pi-voice

Voice input for Pi. Multi-provider STT with Deepgram streaming, Groq Whisper, and OpenAI Whisper. 56+ languages.

Install

pi install npm:@artale/pi-voice

Setup

Set at least one API key:

# Best quality — Deepgram Nova 3 with streaming ($200 free credit)
export DEEPGRAM_API_KEY="your-key"

# Fast and free — Groq Whisper
export GROQ_API_KEY="your-key"

# Reliable — OpenAI Whisper
export OPENAI_API_KEY="your-key"

Install an audio capture tool:

brew install sox        # macOS/Linux (recommended)
brew install ffmpeg     # alternative
# arecord is pre-installed on most Linux distros

Verify: /voice test

Commands

/voice [seconds]        — record and transcribe (default 10s)
/voice dictate          — continuous dictation (30s)
/voice test             — check audio + provider setup
/voice config [prov]    — set provider (auto/deepgram/groq/openai)
/voice lang [code]      — set language (en, es, fr, de, ja, zh, ko, ...)
/voice history          — recent transcriptions
/voice stop             — stop active recording

Provider fallback chain

1. Deepgram Nova 3 — streaming via WebSocket, best quality, 56+ languages
2. Groq Whisper Large v3 Turbo — fast, free tier
3. OpenAI Whisper-1 — reliable, paid

Auto-detection: uses the first available provider. Override with /voice config deepgram.

56+ Languages

/voice lang es          — Spanish
/voice lang ja          — Japanese
/voice lang zh          — Chinese (auto-switches to Nova 2)
/voice lang de          — German
/voice lang fr          — French

Full list: /voice lang

What's new in v2.0

• Deepgram streaming via WebSocket + REST fallback
• Multi-provider fallback chain (Deepgram → Groq → OpenAI)
• 56+ languages with auto model selection
• Continuous dictation mode (/voice dictate)
• Provider diagnostics (/voice test)
• Better Windows support (ffmpeg dshow)
• History with provider and language tracking

License

MIT